README.md

metadata

title: Falcon-Perception-0.6B WebGPU
emoji: 🦅
colorFrom: indigo
colorTo: pink
sdk: static
pinned: false
license: apache-2.0
short_description: Open-vocab detection + segmentation, all in the browser
models:
  - tiiuae/Falcon-Perception
  - onnx-community/falcon-perception-onnx-webgpu

🦅 Falcon-Perception-0.6B WebGPU

A browser demo for tiiuae/Falcon-Perception — a 0.6B open-vocabulary VLM that turns natural-language queries into bounding boxes and pixel-accurate segmentation masks, running fully client-side via WebGPU + ONNX Runtime Web.

What's inside

• Detection — draw bounding boxes for any natural-language query ("athletes", "the runner in front", "mangoes").
• Segmentation — pixel-accurate masks via the AnyUp upsampler, all in-browser.
• Tracker (preview) — HUD-style reticles on video. Limited by VLM latency between detections; see the in-space disclaimer.

How it runs

2.4 GB of ONNX weights are fetched once on first visit, then cached by your browser — no backend, no API keys, no network round-trip after load. Multi-threaded WASM is enabled via coi-serviceworker.